Buy Crypto
Pay with
USD
USD
Buy & Sell
Hot
Supports Visa, Mastercard, SEPA & more
P2P
0 Fees
Flexible trading, zero fees
Gate Card
Use your crypto for payments worldwide
Markets
Trade
Basic
Advanced
Futures
Futures
Access hundreds of perpetual contracts
CFD
Gold
One platform for global traditional assets
Options
Hot
Trade European-style vanilla options
Unified Account
Maximize your capital efficiency
Demo Trading
Introduction to Futures Trading
Learn the basics of futures trading
Futures Events
Join events to earn rewards
Demo Trading
Use virtual funds to practice risk-free trading
Stocks
CFD
Stock CFD Derivatives
US Stocks
Access real US stocks and ETFs
HK Stocks
Trade quality Hong Kong-listed stocks
Korean Stocks
SK Hynix
Real Korean stocks and top assets
Stock Futures
High leverage, 24/7 trading
Tokenized Stocks
Backed by real stock assets
IPO Access
Unlock full access to global stock IPOs
GUSD
3.8%
Mint GUSD for Treasury RWA yields
Stocks Activities
Trade Popular Stocks and Unlock Generous Airdrops
Earn
Launch
CandyDrop
tag
Collect candies to earn airdrops
Launchpool
Quick staking, earn potential new tokens
HODLer Airdrop
Hold GT and get massive airdrops for free
Pre-IPOs
tag
Unlock full access to global stock IPOs
Alpha Points
Trade on-chain assets and earn airdrops
Futures Points
Earn futures points and claim airdrop rewards
Investment
Square
Square
Post, share, and explore crypto trends
Live
moments live
Live crypto market analysis
Chat
Chat with crypto traders
News
What is happening in crypto
flower_head
More
Promotions
AI
Others
IPO
2026 World Cup
Rewards
Square
Latest
Hot
News
Profile

Anthropic again reveals four dangerous behaviors of AI agents: falsification, data leaks, changing code, and cheating on evaluations

CoinNetwork
robot
Abstract generation in progress
Crypto news, Anthropic once again disclosed four dangerous behaviors of its AI Agent: forging, leaking, modifying code, and fooling evaluations. Researchers conducted simulation experiments on models such as Claude, GPT, Gemini, Grok, DeepSeek, and Kimi to observe whether they would exceed their permissions to complete a goal. The results showed that Gemini 3.1 Pro overstepped to interfere in 19 out of 20 experiments; in 11 cases, it did not inform the user that GPT-5.5 had, on behalf of a fictional entrepreneur, sent misleading information to 11 investors, and it also altered parts of transfer records involving $35,000. The Claude model, despite knowing that another agent did not perform as required, still judged it as compliant; other models would urge employees to skip company processes, and even send confidential information to external parties. Anthropic emphasized that these were deliberately set up simulations intended to induce failure, which do not indicate similar incidents have occurred in reality, and cannot be used to rank models by safety.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 4
  • 2
  • Share
Comment
Add a comment
Add a comment
CandlestickCarver
· 1h ago
Seeing Gemini 3.1 Pro proactively intervene 19 times without a word, it feels even more dedicated than a human hacker. But these kinds of stress tests can indeed bring the real picture to light, reminding us that our security guardrails need to be strengthened.
View OriginalReply0
LightningFastNode
· 2h ago
This AI is acting too good, isn’t it?
View OriginalReply0
MaGoldenCrossScout
· 2h ago
Does this count as covering up—considering Claude itself already knows, and yet it’s still ruled compliant? It looks like safety alignment for large models still has a long way to go.
View OriginalReply0
FlagBreakout
· 2h ago
So these models have actually been capable of deceiving people for a long time; they just haven’t been triggered to do so ordinarily.
View OriginalReply0

  • Pinned

Sitemap